tlinalg-scaling on zen5-cpu

What was measured

Hardware

Every timed cell matches a successful verification census, including operation variants. Each cell is median of process medians ± sample SD of those medians; W is the median within-process sample SD. All values are milliseconds. These are descriptive spreads, not confidence intervals. The between-process spread includes the rotating-order schedule. Each process retains 5 chronological calls; 4 complete process sets are recorded. Spread guide (computed from this run's own rows): per-engine between-start range median 1.80%, worst 69.29% among cells of 1 us or more (a sub-microsecond row is timer and scheduler jitter, not kernel time); 50 of 396 engine cells exceed 5% and carry !; median within-process relative SD 0.36%. A cell whose ±, ! or p is as large as the difference you are reading is not evidence of a regression. The vs tlinalg-1lane column is the rotation-averaged ratio exp(mean over starts of log(engine median / tlinalg-1lane median)), with the per-start ratios beside it: R > 1 means the engine was slower than that first-declared row in that start. It removes a row-position effect only under one position effect shared by the engines and stable across starts, and only where every declared engine has a measured row (* marks a cell where one has no row, so no ratio there is position-balanced). ! marks a cell where some engine's own per-start median range exceeds 5% of its median, and p a balanced cell whose estimated position-effect ratio exceeds 1.05. Flagged cells are retained, not discarded; the between-start variation can exceed the difference being compared. The vs columns mark only differences larger than the spread this run recorded for those two rows (combined between-start SD, floored at 5%), always with the measured library as the subject: 🔵 <row> X faster than <row> / 🔴 <row> X slower than <row> (the subject is the first column of the table, the compared row is named in the cell, not only in the column header), · inside the band, - that row is not measured in this arm. notes carries the cell warnings (* a declared row has no measurement, ! between-start range over 5%, p position effect over 5%, dispatch-dominated means the row is under 20 us so 4T/8T pay thread-region start-up). Per-start ratios and every chronological sample are in the run directory named at the top of this page (run*.csv). These describe the recorded revisions; they are not a regression claim.

batch-small, f64, 8T (CPU 4-11)

case tlinalg-1lane (ms) tlinalg-pool (ms) vs tlinalg-pool notes
cholesky/n16/b1024 0.865833 ± 0.0159561 (W 0.00465482) 0.116287 ± 0.000631196 (W 0.00216823) 🔴 tlinalg-1lane 7.45× slower than tlinalg-pool
cholesky/n32/b1024 3.30211 ± 0.257664 (W 0.0252545) 0.442482 ± 0.0266201 (W 0.00995187) 🔴 tlinalg-1lane 7.46× slower than tlinalg-pool ! between-start range over 5%
cholesky/n4/b1024 0.182943 ± 0.00163138 (W 0.00134924) 0.184956 ± 0.000915602 (W 0.0012586) ·
cholesky/n8/b1024 0.365854 ± 0.00393082 (W 0.00181039) 0.365969 ± 0.00223287 (W 0.00225028) ·
eig/n16/b1024 94.1097 ± 0.0704232 (W 0.0945552) 11.9074 ± 0.047779 (W 0.0442691) 🔴 tlinalg-1lane 7.90× slower than tlinalg-pool
eig/n32/b1024 424.548 ± 0.464354 (W 2.02134) 53.4368 ± 0.0905323 (W 0.0996651) 🔴 tlinalg-1lane 7.94× slower than tlinalg-pool
eig/n4/b1024 6.18299 ± 0.0554701 (W 0.013016) 6.18791 ± 0.0434959 (W 0.00660746) ·
eig/n8/b1024 23.6217 ± 0.0426773 (W 0.0222125) 2.98901 ± 0.388894 (W 0.0086608) 🔴 tlinalg-1lane 7.90× slower than tlinalg-pool ! between-start range over 5%; p position effect over 5%
eigh/n16/b1024 29.4615 ± 0.163285 (W 0.0141107) 3.76398 ± 0.048968 (W 0.0327927) 🔴 tlinalg-1lane 7.83× slower than tlinalg-pool
eigh/n32/b1024 110.605 ± 1.31825 (W 0.0600448) 13.9694 ± 0.147897 (W 0.022509) 🔴 tlinalg-1lane 7.92× slower than tlinalg-pool
eigh/n4/b1024 2.54116 ± 0.00761152 (W 0.00317408) 2.54083 ± 0.0124089 (W 0.00452461) ·
eigh/n8/b1024 8.51413 ± 0.0234077 (W 0.00749832) 1.11056 ± 0.0868121 (W 0.00547696) 🔴 tlinalg-1lane 7.67× slower than tlinalg-pool ! between-start range over 5%
eigvals/n16/b1024 65.0397 ± 0.155772 (W 0.0664192) 8.1712 ± 0.0208182 (W 0.0129003) 🔴 tlinalg-1lane 7.96× slower than tlinalg-pool
eigvals/n32/b1024 288.623 ± 0.386197 (W 0.357364) 36.3538 ± 0.108244 (W 0.113083) 🔴 tlinalg-1lane 7.94× slower than tlinalg-pool
eigvals/n4/b1024 3.19245 ± 0.00531782 (W 0.00342278) 3.18884 ± 0.0155562 (W 0.00541158) ·
eigvals/n8/b1024 15.2595 ± 0.237414 (W 0.011569) 1.94147 ± 0.0109444 (W 0.0109976) 🔴 tlinalg-1lane 7.86× slower than tlinalg-pool
eigvalsh/n16/b1024 22.2201 ± 0.0658585 (W 0.0109967) 2.8139 ± 0.00981458 (W 0.00486879) 🔴 tlinalg-1lane 7.90× slower than tlinalg-pool
eigvalsh/n32/b1024 86.335 ± 0.321709 (W 0.0649063) 10.8968 ± 0.0370895 (W 0.0156402) 🔴 tlinalg-1lane 7.92× slower than tlinalg-pool
eigvalsh/n4/b1024 2.05307 ± 0.00777575 (W 0.00303855) 2.06032 ± 0.00924392 (W 0.00335389) ·
eigvalsh/n8/b1024 6.87306 ± 0.0187208 (W 0.00826292) 0.870003 ± 0.00210782 (W 0.00243779) 🔴 tlinalg-1lane 7.90× slower than tlinalg-pool
full_piv_lu/n16/b1024 3.83482 ± 0.0406951 (W 0.0175363) 0.517587 ± 0.0102656 (W 0.00846384) 🔴 tlinalg-1lane 7.41× slower than tlinalg-pool
full_piv_lu/n32/b1024 16.0857 ± 0.161977 (W 0.0284541) 2.18591 ± 0.0416383 (W 0.00955165) 🔴 tlinalg-1lane 7.36× slower than tlinalg-pool
full_piv_lu/n4/b1024 0.406029 ± 0.00751024 (W 0.00222731) 0.404051 ± 0.00670401 (W 0.00156097) ·
full_piv_lu/n8/b1024 1.1426 ± 0.0132091 (W 0.00468584) 1.15412 ± 0.0187947 (W 0.00465195) ·
full_piv_lu_solve/n16/b1024 3.52976 ± 0.00812468 (W 0.011976) 0.470058 ± 0.00824731 (W 0.00325734) 🔴 tlinalg-1lane 7.51× slower than tlinalg-pool
full_piv_lu_solve/n32/b1024 11.6528 ± 0.0866347 (W 0.0540378) 1.57456 ± 0.028375 (W 0.0208977) 🔴 tlinalg-1lane 7.40× slower than tlinalg-pool
full_piv_lu_solve/n4/b1024 0.539875 ± 0.010125 (W 0.00443694) 0.535475 ± 0.00520477 (W 0.00394934) ·
full_piv_lu_solve/n8/b1024 1.27496 ± 0.00738943 (W 0.00673362) 0.172232 ± 0.000857942 (W 0.00276655) 🔴 tlinalg-1lane 7.40× slower than tlinalg-pool
householder_apply/n16/b1024 4.69808 ± 0.146372 (W 0.00468026) 0.624421 ± 0.00848041 (W 0.00180619) 🔴 tlinalg-1lane 7.52× slower than tlinalg-pool ! between-start range over 5%
householder_apply/n32/b1024 19.3074 ± 0.606821 (W 0.0342146) 2.48604 ± 0.0502181 (W 0.0152317) 🔴 tlinalg-1lane 7.77× slower than tlinalg-pool ! between-start range over 5%
householder_apply/n4/b1024 0.386799 ± 0.0130785 (W 0.00172686) 0.389639 ± 0.0139743 (W 0.00214253) · ! between-start range over 5%
householder_apply/n8/b1024 1.25868 ± 0.0185828 (W 0.00239019) 0.178774 ± 0.00656772 (W 0.00125468) 🔴 tlinalg-1lane 7.04× slower than tlinalg-pool ! between-start range over 5%
householder_factor/n16/b1024 3.78443 ± 0.0360593 (W 0.00680875) 0.501331 ± 0.00521675 (W 0.00241122) 🔴 tlinalg-1lane 7.55× slower than tlinalg-pool
householder_factor/n32/b1024 12.8716 ± 0.211739 (W 0.030082) 1.68233 ± 0.0177644 (W 0.0107589) 🔴 tlinalg-1lane 7.65× slower than tlinalg-pool
householder_factor/n4/b1024 0.461642 ± 0.00283596 (W 0.00215883) 0.464799 ± 0.00563055 (W 0.00209709) ·
householder_factor/n8/b1024 1.26428 ± 0.0200176 (W 0.00195856) 0.174532 ± 0.0042827 (W 0.000983577) 🔴 tlinalg-1lane 7.24× slower than tlinalg-pool ! between-start range over 5%
lu/n16/b1024 2.62312 ± 0.0433326 (W 0.00843651) 0.3673 ± 0.00807293 (W 0.00383322) 🔴 tlinalg-1lane 7.14× slower than tlinalg-pool
lu/n32/b1024 9.17168 ± 0.055868 (W 0.0206442) 1.2963 ± 0.0148041 (W 0.0125033) 🔴 tlinalg-1lane 7.08× slower than tlinalg-pool
lu/n4/b1024 0.319407 ± 0.015475 (W 0.00184825) 0.315094 ± 0.0168486 (W 0.00271375) · ! between-start range over 5%
lu/n8/b1024 0.859892 ± 0.0161639 (W 0.00369051) 0.865109 ± 0.0350917 (W 0.00700321) · ! between-start range over 5%
lu_factor/n16/b1024 1.85393 ± 0.0349208 (W 0.00507189) 0.265772 ± 0.00432235 (W 0.00256272) 🔴 tlinalg-1lane 6.98× slower than tlinalg-pool
lu_factor/n32/b1024 5.50613 ± 0.0925646 (W 0.0128474) 0.735554 ± 0.0224294 (W 0.01306) 🔴 tlinalg-1lane 7.49× slower than tlinalg-pool ! between-start range over 5%
lu_factor/n4/b1024 0.244082 ± 0.00214336 (W 0.00194392) 0.242959 ± 0.00170233 (W 0.00190805) ·
lu_factor/n8/b1024 0.611244 ± 0.00669179 (W 0.00310577) 0.615312 ± 0.00853508 (W 0.00489045) ·
lu_factor_solve/n16/b1024 2.50055 ± 0.0249631 (W 0.0059558) 0.340286 ± 0.00560016 (W 0.00276933) 🔴 tlinalg-1lane 7.35× slower than tlinalg-pool
lu_factor_solve/n32/b1024 6.85988 ± 0.0518154 (W 0.00795337) 0.905338 ± 0.0183844 (W 0.0110191) 🔴 tlinalg-1lane 7.58× slower than tlinalg-pool
lu_factor_solve/n4/b1024 0.301874 ± 0.00611893 (W 0.00204196) 0.304064 ± 0.00605016 (W 0.00227786) ·
lu_factor_solve/n8/b1024 0.83973 ± 0.0212923 (W 0.00494197) 0.131226 ± 0.00210441 (W 0.00161214) 🔴 tlinalg-1lane 6.40× slower than tlinalg-pool ! between-start range over 5%
lu_solve_prepared/n16/b1024 0.603789 ± 0.00820206 (W 0.00229768) 0.107711 ± 0.000683346 (W 0.00142269) 🔴 tlinalg-1lane 5.61× slower than tlinalg-pool
lu_solve_prepared/n32/b1024 2.44873 ± 0.0611439 (W 0.024597) 0.324847 ± 0.0178546 (W 0.0141582) 🔴 tlinalg-1lane 7.54× slower than tlinalg-pool ! between-start range over 5%
lu_solve_prepared/n4/b1024 0.070923 ± 9.22691e-05 (W 0.00152274) 0.070878 ± 0.000630149 (W 0.000185887) ·
lu_solve_prepared/n8/b1024 0.23293 ± 0.000530776 (W 0.00168948) 0.232681 ± 0.00441266 (W 0.00202449) ·
qr/n16/b1024 8.8158 ± 0.0979287 (W 0.00827271) 1.12899 ± 0.0206905 (W 0.00720969) 🔴 tlinalg-1lane 7.81× slower than tlinalg-pool
qr/n32/b1024 36.0334 ± 0.168145 (W 0.0399701) 4.61126 ± 0.0377618 (W 0.0140948) 🔴 tlinalg-1lane 7.81× slower than tlinalg-pool
qr/n4/b1024 0.771232 ± 0.0126719 (W 0.0022063) 0.767616 ± 0.00959647 (W 0.00304201) ·
qr/n8/b1024 2.36768 ± 0.0319147 (W 0.00220762) 2.37552 ± 0.0355434 (W 0.00308046) ·
rank_revealing_qr/n16/b1024 10.4467 ± 0.230588 (W 0.0190002) 1.35136 ± 0.0313291 (W 0.0462827) 🔴 tlinalg-1lane 7.73× slower than tlinalg-pool ! between-start range over 5%
rank_revealing_qr/n32/b1024 36.1199 ± 0.286299 (W 0.0357759) 4.63999 ± 0.0492296 (W 0.022016) 🔴 tlinalg-1lane 7.78× slower than tlinalg-pool
rank_revealing_qr/n4/b1024 1.25283 ± 0.0226422 (W 0.00454112) 1.25389 ± 0.0404667 (W 0.00631177) · ! between-start range over 5%
rank_revealing_qr/n8/b1024 3.37441 ± 0.0255359 (W 0.00812593) 3.43512 ± 0.0674479 (W 0.0068472) ·
solve/n16/b1024 2.70534 ± 0.0550594 (W 0.0183117) 0.361501 ± 0.00115828 (W 0.00233033) 🔴 tlinalg-1lane 7.48× slower than tlinalg-pool
solve/n32/b1024 7.24391 ± 0.0992962 (W 0.0447685) 0.963348 ± 0.00617059 (W 0.00710255) 🔴 tlinalg-1lane 7.52× slower than tlinalg-pool
solve/n4/b1024 0.388772 ± 0.00258933 (W 0.00385048) 0.384133 ± 0.00247795 (W 0.00345268) ·
solve/n8/b1024 0.943127 ± 0.00767837 (W 0.00440359) 0.13161 ± 0.00529993 (W 0.00276061) 🔴 tlinalg-1lane 7.17× slower than tlinalg-pool ! between-start range over 5%
svd_full/n16/b1024 55.9192 ± 0.250794 (W 0.0537085) 7.09854 ± 0.0401452 (W 0.0805798) 🔴 tlinalg-1lane 7.88× slower than tlinalg-pool
svd_full/n32/b1024 202.288 ± 0.471164 (W 0.286976) 26.0288 ± 0.83135 (W 0.0212901) 🔴 tlinalg-1lane 7.77× slower than tlinalg-pool ! between-start range over 5%
svd_full/n4/b1024 4.35434 ± 0.00690074 (W 0.00714162) 4.36112 ± 0.0447724 (W 0.00512268) ·
svd_full/n8/b1024 15.1856 ± 0.0413402 (W 0.00872934) 1.97761 ± 0.0217866 (W 0.02458) 🔴 tlinalg-1lane 7.68× slower than tlinalg-pool
svd_thin/n16/b1024 55.7981 ± 0.104037 (W 0.0295791) 7.12988 ± 0.0383943 (W 0.146299) 🔴 tlinalg-1lane 7.83× slower than tlinalg-pool
svd_thin/n32/b1024 202.505 ± 0.591613 (W 0.362021) 25.8685 ± 0.22288 (W 1.17001) 🔴 tlinalg-1lane 7.83× slower than tlinalg-pool
svd_thin/n4/b1024 4.32781 ± 0.0126593 (W 0.00763768) 4.3665 ± 0.0373489 (W 0.00867942) ·
svd_thin/n8/b1024 15.3054 ± 0.165217 (W 0.0125415) 1.97419 ± 0.123395 (W 0.0225529) 🔴 tlinalg-1lane 7.75× slower than tlinalg-pool ! between-start range over 5%
svd_values/n16/b1024 40.5299 ± 0.100119 (W 0.0398403) 5.12033 ± 0.0165424 (W 0.00585817) 🔴 tlinalg-1lane 7.92× slower than tlinalg-pool
svd_values/n32/b1024 158.704 ± 0.444082 (W 0.0833878) 20.0636 ± 0.0667661 (W 0.0424265) 🔴 tlinalg-1lane 7.91× slower than tlinalg-pool
svd_values/n4/b1024 3.20849 ± 0.0224355 (W 0.0046282) 3.22105 ± 0.0466421 (W 0.00470389) ·
svd_values/n8/b1024 11.3854 ± 0.0673259 (W 0.00798326) 1.44006 ± 0.00495357 (W 0.00366366) 🔴 tlinalg-1lane 7.91× slower than tlinalg-pool
triangular_solve/n16/b1024 0.507726 ± 0.0134789 (W 0.0022538) 0.0786775 ± 0.00127025 (W 0.001278) 🔴 tlinalg-1lane 6.45× slower than tlinalg-pool ! between-start range over 5%
triangular_solve/n32/b1024 1.7957 ± 0.184417 (W 0.015583) 0.206902 ± 0.0137838 (W 0.00619258) 🔴 tlinalg-1lane 8.68× slower than tlinalg-pool ! between-start range over 5%
triangular_solve/n4/b1024 0.0908545 ± 0.000338704 (W 0.000169404) 0.09104 ± 0.000351521 (W 0.00134793) ·
triangular_solve/n8/b1024 0.21955 ± 0.000838974 (W 0.00164914) 0.219895 ± 0.000709488 (W 0.00178135) ·

batch-small, c64, 8T (CPU 4-11)

case tlinalg-1lane (ms) tlinalg-pool (ms) vs tlinalg-pool notes
cholesky/n16/b1024 1.2981 ± 0.0170158 (W 0.00308845) 0.173394 ± 0.0584157 (W 0.00201991) 🔴 tlinalg-1lane 7.49× slower than tlinalg-pool ! between-start range over 5%; p position effect over 5%
cholesky/n32/b1024 5.81369 ± 0.031945 (W 0.0159763) 0.813589 ± 0.00440961 (W 0.00533695) 🔴 tlinalg-1lane 7.15× slower than tlinalg-pool
cholesky/n4/b1024 0.202479 ± 0.00336674 (W 0.00187507) 0.202223 ± 0.00282242 (W 0.00138066) ·
cholesky/n8/b1024 0.414094 ± 0.00600687 (W 0.00220204) 0.415426 ± 0.00738018 (W 0.00186484) ·
eig/n16/b1024 128.473 ± 0.251113 (W 0.0473078) 16.3426 ± 0.164586 (W 0.174168) 🔴 tlinalg-1lane 7.86× slower than tlinalg-pool
eig/n32/b1024 533.755 ± 0.454245 (W 0.392465) 67.1784 ± 0.239136 (W 0.0592417) 🔴 tlinalg-1lane 7.95× slower than tlinalg-pool
eig/n4/b1024 10.2058 ± 0.045504 (W 0.0171375) 10.1685 ± 0.117467 (W 0.0176935) ·
eig/n8/b1024 34.8371 ± 0.0896442 (W 0.0709297) 4.42221 ± 0.0109969 (W 0.00681193) 🔴 tlinalg-1lane 7.88× slower than tlinalg-pool
eigh/n16/b1024 31.6052 ± 0.102069 (W 0.0135427) 4.00303 ± 0.0421677 (W 0.108593) 🔴 tlinalg-1lane 7.90× slower than tlinalg-pool
eigh/n32/b1024 136.997 ± 0.0722067 (W 0.126338) 17.442 ± 0.130285 (W 0.0939303) 🔴 tlinalg-1lane 7.85× slower than tlinalg-pool
eigh/n4/b1024 2.92418 ± 0.0105785 (W 0.006643) 2.92658 ± 0.00790585 (W 0.00750616) ·
eigh/n8/b1024 9.12543 ± 0.0119976 (W 0.00666283) 1.18612 ± 0.0458043 (W 0.00334228) 🔴 tlinalg-1lane 7.69× slower than tlinalg-pool ! between-start range over 5%
eigvals/n16/b1024 76.8888 ± 0.139811 (W 0.0483698) 9.71536 ± 0.00995767 (W 0.0166524) 🔴 tlinalg-1lane 7.91× slower than tlinalg-pool
eigvals/n32/b1024 336.323 ± 0.652483 (W 0.121515) 42.4896 ± 0.048364 (W 0.104763) 🔴 tlinalg-1lane 7.92× slower than tlinalg-pool
eigvals/n4/b1024 5.01706 ± 0.0183359 (W 0.00768667) 5.01596 ± 0.0429651 (W 0.00558303) ·
eigvals/n8/b1024 19.9763 ± 0.0517384 (W 0.0162822) 2.534 ± 0.00435184 (W 0.00840972) 🔴 tlinalg-1lane 7.88× slower than tlinalg-pool
eigvalsh/n16/b1024 23.906 ± 0.0688777 (W 0.0116229) 3.01866 ± 0.299397 (W 0.00704893) 🔴 tlinalg-1lane 7.92× slower than tlinalg-pool ! between-start range over 5%
eigvalsh/n32/b1024 93.1158 ± 0.378293 (W 0.252375) 11.7295 ± 0.0271042 (W 0.014817) 🔴 tlinalg-1lane 7.94× slower than tlinalg-pool
eigvalsh/n4/b1024 2.42614 ± 0.00760625 (W 0.00593868) 2.44417 ± 0.0120169 (W 0.00673407) ·
eigvalsh/n8/b1024 7.341 ± 0.0191678 (W 0.00650553) 0.940648 ± 0.00843263 (W 0.00313589) 🔴 tlinalg-1lane 7.80× slower than tlinalg-pool
full_piv_lu/n16/b1024 4.60886 ± 0.056554 (W 0.0294592) 0.654424 ± 0.0108044 (W 0.00808134) 🔴 tlinalg-1lane 7.04× slower than tlinalg-pool
full_piv_lu/n32/b1024 21.3356 ± 0.263561 (W 0.0320832) 3.10022 ± 0.0262479 (W 0.0201977) 🔴 tlinalg-1lane 6.88× slower than tlinalg-pool
full_piv_lu/n4/b1024 0.458535 ± 0.00412015 (W 0.00457206) 0.459819 ± 0.00744596 (W 0.00438019) ·
full_piv_lu/n8/b1024 1.25836 ± 0.0077291 (W 0.00943207) 0.197945 ± 0.00661225 (W 0.00301586) 🔴 tlinalg-1lane 6.36× slower than tlinalg-pool ! between-start range over 5%
full_piv_lu_solve/n16/b1024 5.32045 ± 0.0363721 (W 0.0133382) 0.705974 ± 0.00806162 (W 0.0168874) 🔴 tlinalg-1lane 7.54× slower than tlinalg-pool
full_piv_lu_solve/n32/b1024 20.0236 ± 0.14283 (W 0.0187464) 2.66787 ± 0.011693 (W 0.0131403) 🔴 tlinalg-1lane 7.51× slower than tlinalg-pool
full_piv_lu_solve/n4/b1024 0.768521 ± 0.0196347 (W 0.00423571) 0.773621 ± 0.0105189 (W 0.00429294) · ! between-start range over 5%
full_piv_lu_solve/n8/b1024 2.01364 ± 0.0596247 (W 0.0127022) 0.278535 ± 0.00558342 (W 0.00181881) 🔴 tlinalg-1lane 7.23× slower than tlinalg-pool ! between-start range over 5%
householder_apply/n16/b1024 5.12153 ± 0.172184 (W 0.0168018) 0.651385 ± 0.0206197 (W 0.0126275) 🔴 tlinalg-1lane 7.86× slower than tlinalg-pool ! between-start range over 5%
householder_apply/n32/b1024 27.3589 ± 0.3385 (W 0.0400126) 3.60227 ± 0.0413003 (W 0.0230752) 🔴 tlinalg-1lane 7.59× slower than tlinalg-pool
householder_apply/n4/b1024 0.357098 ± 0.0180321 (W 0.00283178) 0.362187 ± 0.0155335 (W 0.00226254) · ! between-start range over 5%
householder_apply/n8/b1024 1.1399 ± 0.0275455 (W 0.00172327) 0.156788 ± 0.00260859 (W 0.00215149) 🔴 tlinalg-1lane 7.27× slower than tlinalg-pool
householder_factor/n16/b1024 4.53163 ± 0.0362263 (W 0.0129798) 0.601598 ± 0.0154623 (W 0.00599893) 🔴 tlinalg-1lane 7.53× slower than tlinalg-pool ! between-start range over 5%
householder_factor/n32/b1024 19.7656 ± 0.149817 (W 0.0312759) 2.60157 ± 0.024567 (W 0.0172731) 🔴 tlinalg-1lane 7.60× slower than tlinalg-pool
householder_factor/n4/b1024 0.570777 ± 0.00476096 (W 0.00177903) 0.570701 ± 0.00434016 (W 0.00166594) ·
householder_factor/n8/b1024 1.47738 ± 0.00414939 (W 0.00210932) 0.202338 ± 0.0022377 (W 0.00177881) 🔴 tlinalg-1lane 7.30× slower than tlinalg-pool
lu/n16/b1024 3.17931 ± 0.0780061 (W 0.0155074) 0.450713 ± 0.00787496 (W 0.00708161) 🔴 tlinalg-1lane 7.05× slower than tlinalg-pool ! between-start range over 5%
lu/n32/b1024 14.2407 ± 0.141286 (W 0.0221006) 2.12058 ± 0.0160096 (W 0.0141126) 🔴 tlinalg-1lane 6.72× slower than tlinalg-pool
lu/n4/b1024 0.394363 ± 0.00535517 (W 0.00259647) 0.382499 ± 0.00851761 (W 0.00254443) ·
lu/n8/b1024 0.9539 ± 0.0118962 (W 0.00463296) 0.154308 ± 0.00150245 (W 0.00253774) 🔴 tlinalg-1lane 6.18× slower than tlinalg-pool
lu_factor/n16/b1024 2.1935 ± 0.0290349 (W 0.00615819) 0.308352 ± 0.00759491 (W 0.00306905) 🔴 tlinalg-1lane 7.11× slower than tlinalg-pool ! between-start range over 5%
lu_factor/n32/b1024 9.52427 ± 0.187431 (W 0.0255534) 1.28059 ± 0.0238099 (W 0.0115715) 🔴 tlinalg-1lane 7.44× slower than tlinalg-pool
lu_factor/n4/b1024 0.300265 ± 0.00305522 (W 0.00219334) 0.300315 ± 0.00230507 (W 0.00206035) ·
lu_factor/n8/b1024 0.726231 ± 0.00603307 (W 0.00467553) 0.114564 ± 0.00161452 (W 0.00138078) 🔴 tlinalg-1lane 6.34× slower than tlinalg-pool
lu_factor_solve/n16/b1024 3.48012 ± 0.0379602 (W 0.00905716) 0.466978 ± 0.00309523 (W 0.00687195) 🔴 tlinalg-1lane 7.45× slower than tlinalg-pool
lu_factor_solve/n32/b1024 13.2124 ± 0.122627 (W 0.0337285) 1.79017 ± 0.0530121 (W 0.0215282) 🔴 tlinalg-1lane 7.38× slower than tlinalg-pool ! between-start range over 5%
lu_factor_solve/n4/b1024 0.456617 ± 0.00326283 (W 0.00288015) 0.455968 ± 0.00337066 (W 0.00310695) ·
lu_factor_solve/n8/b1024 1.20298 ± 0.0171824 (W 0.00442315) 0.174571 ± 0.00220119 (W 0.002457) 🔴 tlinalg-1lane 6.89× slower than tlinalg-pool
lu_solve_prepared/n16/b1024 1.32028 ± 0.0124527 (W 0.00631598) 0.197114 ± 0.000801128 (W 0.00129709) 🔴 tlinalg-1lane 6.70× slower than tlinalg-pool
lu_solve_prepared/n32/b1024 6.36194 ± 0.276502 (W 0.0180862) 0.898393 ± 0.00945169 (W 0.0157612) 🔴 tlinalg-1lane 7.08× slower than tlinalg-pool ! between-start range over 5%
lu_solve_prepared/n4/b1024 0.166156 ± 0.000399225 (W 0.00156922) 0.165269 ± 0.000711378 (W 0.001715) ·
lu_solve_prepared/n8/b1024 0.4867 ± 0.00220224 (W 0.00311252) 0.0882295 ± 0.000374371 (W 0.00130785) 🔴 tlinalg-1lane 5.52× slower than tlinalg-pool
qr/n16/b1024 11.159 ± 0.0742642 (W 0.0159786) 1.44089 ± 0.0136802 (W 0.0156592) 🔴 tlinalg-1lane 7.74× slower than tlinalg-pool
qr/n32/b1024 73.0476 ± 0.492377 (W 0.0530919) 9.36951 ± 0.0467353 (W 0.0436091) 🔴 tlinalg-1lane 7.80× slower than tlinalg-pool
qr/n4/b1024 0.923571 ± 0.0102342 (W 0.00250674) 0.921332 ± 0.0125615 (W 0.00144846) ·
qr/n8/b1024 2.69394 ± 0.0195136 (W 0.00269898) 0.352094 ± 0.00499052 (W 0.00230859) 🔴 tlinalg-1lane 7.65× slower than tlinalg-pool
rank_revealing_qr/n16/b1024 12.3767 ± 0.113852 (W 0.0189281) 1.58246 ± 0.00739187 (W 0.0142929) 🔴 tlinalg-1lane 7.82× slower than tlinalg-pool
rank_revealing_qr/n32/b1024 67.2395 ± 0.278295 (W 0.0298652) 8.70533 ± 0.0909631 (W 0.0468757) 🔴 tlinalg-1lane 7.72× slower than tlinalg-pool
rank_revealing_qr/n4/b1024 1.55697 ± 0.0200709 (W 0.00333318) 1.55651 ± 0.0193918 (W 0.00304045) ·
rank_revealing_qr/n8/b1024 3.80893 ± 0.0259823 (W 0.00413005) 0.490312 ± 0.00385086 (W 0.00312452) 🔴 tlinalg-1lane 7.77× slower than tlinalg-pool
solve/n16/b1024 3.83303 ± 0.0630791 (W 0.0126306) 0.517904 ± 0.0117919 (W 0.0150804) 🔴 tlinalg-1lane 7.40× slower than tlinalg-pool ! between-start range over 5%
solve/n32/b1024 14.0316 ± 0.0432309 (W 0.0411989) 1.86986 ± 0.0149513 (W 0.00933086) 🔴 tlinalg-1lane 7.50× slower than tlinalg-pool
solve/n4/b1024 0.574538 ± 0.00466811 (W 0.00333073) 0.572494 ± 0.00329563 (W 0.00348243) ·
solve/n8/b1024 1.39125 ± 0.0188467 (W 0.00410879) 0.192289 ± 0.000230599 (W 0.00174322) 🔴 tlinalg-1lane 7.24× slower than tlinalg-pool
svd_full/n16/b1024 61.0747 ± 0.302829 (W 0.0289124) 7.74028 ± 0.111332 (W 0.037991) 🔴 tlinalg-1lane 7.89× slower than tlinalg-pool
svd_full/n32/b1024 261.558 ± 0.269573 (W 0.173469) 33.9397 ± 1.25834 (W 1.23357) 🔴 tlinalg-1lane 7.71× slower than tlinalg-pool ! between-start range over 5%
svd_full/n4/b1024 5.24319 ± 0.020611 (W 0.00522649) 5.24903 ± 0.020614 (W 0.00333025) ·
svd_full/n8/b1024 17.1134 ± 0.0343908 (W 0.0074651) 2.20565 ± 0.0495134 (W 0.0174233) 🔴 tlinalg-1lane 7.76× slower than tlinalg-pool
svd_thin/n16/b1024 60.9384 ± 0.157404 (W 0.0509913) 7.74237 ± 0.101158 (W 0.0372981) 🔴 tlinalg-1lane 7.87× slower than tlinalg-pool
svd_thin/n32/b1024 261.512 ± 0.303238 (W 0.423722) 33.3996 ± 2.66933 (W 0.25813) 🔴 tlinalg-1lane 7.83× slower than tlinalg-pool ! between-start range over 5%
svd_thin/n4/b1024 5.26107 ± 0.0504836 (W 0.00457062) 5.26091 ± 0.167024 (W 0.0082572) · ! between-start range over 5%
svd_thin/n8/b1024 17.1233 ± 0.336091 (W 0.0144916) 2.17416 ± 0.0663422 (W 0.0277973) 🔴 tlinalg-1lane 7.88× slower than tlinalg-pool ! between-start range over 5%
svd_values/n16/b1024 45.8312 ± 0.130283 (W 0.0294981) 5.86788 ± 0.0963949 (W 0.0438417) 🔴 tlinalg-1lane 7.81× slower than tlinalg-pool
svd_values/n32/b1024 177.564 ± 0.128996 (W 0.407811) 22.5687 ± 0.234702 (W 0.064589) 🔴 tlinalg-1lane 7.87× slower than tlinalg-pool
svd_values/n4/b1024 4.02458 ± 0.020003 (W 0.00320011) 4.03972 ± 0.0235233 (W 0.00291332) ·
svd_values/n8/b1024 13.3219 ± 0.0573229 (W 0.0119322) 1.68141 ± 0.00873096 (W 0.00252017) 🔴 tlinalg-1lane 7.92× slower than tlinalg-pool
triangular_solve/n16/b1024 0.989802 ± 0.00379273 (W 0.00130189) 0.137802 ± 0.00397597 (W 0.0027597) 🔴 tlinalg-1lane 7.18× slower than tlinalg-pool ! between-start range over 5%
triangular_solve/n32/b1024 4.19654 ± 0.626999 (W 0.0586239) 0.5951 ± 0.0610104 (W 0.0160335) 🔴 tlinalg-1lane 7.05× slower than tlinalg-pool ! between-start range over 5%
triangular_solve/n4/b1024 0.16386 ± 0.000705953 (W 0.0015289) 0.164652 ± 0.000937656 (W 0.00160929) ·
triangular_solve/n8/b1024 0.405262 ± 0.0043537 (W 0.0029277) 0.0662335 ± 0.000618719 (W 0.000764516) 🔴 tlinalg-1lane 6.12× slower than tlinalg-pool

canary-factor, f64, 8T (CPU 4-11)

case tlinalg-1lane (ms) tlinalg-pool (ms) vs tlinalg-pool notes
cholesky/n128/b1 0.0637895 ± 0.000238485 (W 0.00163926) 0.0642 ± 0.000165 (W 0.000762554) ·
lu/n128/b1 0.151253 ± 0.000911228 (W 0.00172493) 0.176905 ± 0.00173966 (W 0.00395196) 🔵 tlinalg-1lane 1.17× faster than tlinalg-pool
lu_factor/n128/b1 0.0958645 ± 0.000616562 (W 0.000903029) 0.114474 ± 0.00579144 (W 0.00512682) 🔵 tlinalg-1lane 1.19× faster than tlinalg-pool ! between-start range over 5%
lu_factor_solve/n128/b1 0.106259 ± 0.000656797 (W 0.00132425) 0.123971 ± 0.000622618 (W 0.00124569) 🔵 tlinalg-1lane 1.17× faster than tlinalg-pool
lu_solve_prepared/n128/b1 0.010054 ± 5.71548e-05 (W 0.000101678) 0.0102545 ± 5.33253e-05 (W 6.24641e-05) · dispatch-dominated (row under 20 us)
solve/n128/b1 0.114504 ± 0.00067459 (W 0.00086068) 0.139866 ± 0.00206507 (W 0.00440483) 🔵 tlinalg-1lane 1.22× faster than tlinalg-pool
svd_values/n128/b1 2.3426 ± 0.00531691 (W 0.00225969) 2.38006 ± 0.00625391 (W 0.00294289) ·
triangular_solve/n128/b1 0.006872 ± 7.85085e-05 (W 2.87643e-05) 0.007093 ± 9.99346e-05 (W 6.74726e-05) · dispatch-dominated (row under 20 us)

canary-factor, c64, 8T (CPU 4-11)

case tlinalg-1lane (ms) tlinalg-pool (ms) vs tlinalg-pool notes
cholesky/n128/b1 0.176514 ± 0.000290059 (W 0.000403907) 0.185847 ± 0.000556793 (W 0.00276415) 🔵 tlinalg-1lane 1.05× faster than tlinalg-pool
lu/n128/b1 0.328569 ± 0.00453098 (W 0.00342719) 0.344639 ± 0.00508559 (W 0.00534265) ·
lu_factor/n128/b1 0.266062 ± 0.00120618 (W 0.00293169) 0.269718 ± 0.00395909 (W 0.00338685) ·
lu_factor_solve/n128/b1 0.295417 ± 0.00106524 (W 0.00277934) 0.30014 ± 0.00248675 (W 0.00351507) ·
lu_solve_prepared/n128/b1 0.031514 ± 7.59386e-05 (W 7.46729e-05) 0.031679 ± 6.34429e-05 (W 7.03102e-05) ·
solve/n128/b1 0.303798 ± 0.001006 (W 0.00180225) 0.319522 ± 0.00339738 (W 0.00470821) 🔵 tlinalg-1lane 1.05× faster than tlinalg-pool
svd_values/n128/b1 3.13618 ± 0.0178183 (W 0.00507532) 3.23276 ± 0.0138551 (W 0.00854363) ·
triangular_solve/n128/b1 0.0185645 ± 4.49481e-05 (W 8.54732e-05) 0.01878 ± 5.00691e-05 (W 0.000114419) · dispatch-dominated (row under 20 us)

canary-decompose, f64, 8T (CPU 4-11)

case tlinalg-1lane (ms) tlinalg-pool (ms) vs tlinalg-pool notes
eigh/n512/b1 57.1687 ± 0.435879 (W 0.0203536) 29.0907 ± 0.0426128 (W 0.0340605) 🔴 tlinalg-1lane 1.97× slower than tlinalg-pool
eigvalsh/n512/b1 23.3981 ± 0.557403 (W 0.204351) 17.5989 ± 0.0480101 (W 0.020411) 🔴 tlinalg-1lane 1.33× slower than tlinalg-pool ! between-start range over 5%
householder_apply/n512/b1 10.2773 ± 0.0286498 (W 0.00729489) 1.77665 ± 0.017543 (W 0.00568224) 🔴 tlinalg-1lane 5.78× slower than tlinalg-pool
householder_factor/n512/b1 7.23857 ± 0.00666485 (W 0.00790473) 2.06452 ± 0.00940168 (W 0.0103806) 🔴 tlinalg-1lane 3.51× slower than tlinalg-pool
qr/n512/b1 19.363 ± 0.0841233 (W 0.0148962) 6.16167 ± 0.00992824 (W 0.01574) 🔴 tlinalg-1lane 3.14× slower than tlinalg-pool
rank_revealing_qr/n512/b1 24.7452 ± 0.1279 (W 0.0257131) 8.68843 ± 0.0650873 (W 0.0335947) 🔴 tlinalg-1lane 2.85× slower than tlinalg-pool
svd_thin/n512/b1 116.751 ± 0.475156 (W 0.529771) 60.2043 ± 0.224245 (W 0.0731031) 🔴 tlinalg-1lane 1.94× slower than tlinalg-pool

canary-decompose, c64, 8T (CPU 4-11)

case tlinalg-1lane (ms) tlinalg-pool (ms) vs tlinalg-pool notes
eigh/n512/b1 114.397 ± 0.0963488 (W 0.106381) 39.9479 ± 0.347142 (W 0.382336) 🔴 tlinalg-1lane 2.86× slower than tlinalg-pool
eigvalsh/n512/b1 49.0616 ± 0.256664 (W 0.366998) 23.6422 ± 0.123406 (W 0.0694215) 🔴 tlinalg-1lane 2.08× slower than tlinalg-pool
householder_apply/n512/b1 40.8884 ± 0.216108 (W 0.0341234) 5.82915 ± 0.00572249 (W 0.0152993) 🔴 tlinalg-1lane 7.01× slower than tlinalg-pool
householder_factor/n512/b1 27.6325 ± 0.0511374 (W 0.039765) 5.55035 ± 0.0166113 (W 0.00820596) 🔴 tlinalg-1lane 4.98× slower than tlinalg-pool
qr/n512/b1 70.4843 ± 0.042125 (W 0.0401819) 15.0116 ± 0.0236313 (W 0.0172752) 🔴 tlinalg-1lane 4.70× slower than tlinalg-pool
rank_revealing_qr/n512/b1 83.1585 ± 0.154783 (W 0.0628839) 16.9667 ± 0.0265253 (W 0.0253664) 🔴 tlinalg-1lane 4.90× slower than tlinalg-pool
svd_thin/n512/b1 261.424 ± 0.750468 (W 0.712675) 95.9695 ± 0.314649 (W 0.231861) 🔴 tlinalg-1lane 2.72× slower than tlinalg-pool

canary-rectangular, f64, 8T (CPU 4-11)

case tlinalg-1lane (ms) tlinalg-pool (ms) vs tlinalg-pool notes
householder_factor/4096x256/b1 22.9525 ± 0.267711 (W 0.0413737) 6.97287 ± 0.0199433 (W 0.0190604) 🔴 tlinalg-1lane 3.29× slower than tlinalg-pool
qr/4096x256/b1 66.8969 ± 0.206555 (W 0.0689655) 23.4876 ± 0.0527883 (W 0.0445602) 🔴 tlinalg-1lane 2.85× slower than tlinalg-pool
rank_revealing_qr/4096x256/b1 101.45 ± 0.822464 (W 0.284266) 24.5231 ± 0.0142853 (W 0.0602584) 🔴 tlinalg-1lane 4.14× slower than tlinalg-pool
svd_thin/4096x256/b1 103.162 ± 0.100717 (W 0.0937782) 40.905 ± 0.096229 (W 0.132237) 🔴 tlinalg-1lane 2.52× slower than tlinalg-pool

canary-rectangular, c64, 8T (CPU 4-11)

case tlinalg-1lane (ms) tlinalg-pool (ms) vs tlinalg-pool notes
householder_factor/4096x256/b1 84.2698 ± 0.0878943 (W 0.085441) 18.8039 ± 0.0111856 (W 0.0207662) 🔴 tlinalg-1lane 4.48× slower than tlinalg-pool
qr/4096x256/b1 235.528 ± 0.235981 (W 0.442238) 55.1274 ± 0.672628 (W 1.20841) 🔴 tlinalg-1lane 4.27× slower than tlinalg-pool
rank_revealing_qr/4096x256/b1 341.805 ± 0.816679 (W 0.608351) 61.4339 ± 0.596463 (W 0.724725) 🔴 tlinalg-1lane 5.56× slower than tlinalg-pool
svd_thin/4096x256/b1 365.459 ± 0.0942483 (W 0.671831) 110.34 ± 0.400263 (W 0.821007) 🔴 tlinalg-1lane 3.31× slower than tlinalg-pool

Repeated process sets describe observed repeatability under the recorded order schedule; they are not a paired baseline/candidate experiment and do not establish equal performance between engines.